# SCAIL 2 Motion Transfer

> SCAIL (Studio-grade Character Animation via In-context Learning) is a character animation model from zai-org that drives a character image with 3D-consistent poses taken from a video. SCAIL 2, the default engine in Kenerate AI's Motion Transfer tool, animates your character or replaces a person in clips up to 2 minutes, and keeps the sound.

URL: https://kenerateai.com/scail-ai
Publisher: Kenerate AI (https://kenerateai.com)
Last updated: 3 October 2026

## Key facts
- Default engine: SCAIL 2
- Modes: Animate character · Replace person
- Driving clip: Up to 2 min
- Output: 480p or 720p
- Sound: Kept on SCAIL 2
- Research: CVPR 2026 Findings
- SCAIL 2 (WaveSpeed): Image + video up to 2 min · Animate or Replace → 480p or 720p · keeps sound. Long clips with sound — animate on your photo's background, or replace the person in the clip.
- SCAIL (WaveSpeed): Image + video up to 2 min → 480p or 720p · silent. The first SCAIL — animate only, silent output, same 2-minute limit.
- Wan 2.2 Animate (Alibaba): Image + video up to 2 min → 480p or 720p · silent. A second opinion on Replace person when SCAIL 2's swap isn't right.
- SteadyDancer (WaveSpeed): Image + video up to 2 min → 480p or 720p · silent. Very fast spins and footwork that need extra stability.
- Kling 3.0 Pro Motion (Kuaishou): Image + 3–30 s video → Follows the source · sound switch. Short clips where face and hand detail matter most.

## Steps
1. Open Motion Transfer — SCAIL 2 is already selected — it's the default engine. Try the 'Street dance' starter to see the settings.
2. Add the character and the clip — A clear JPG or PNG of the character, and a driving video up to 2 minutes (MP4, MOV, WEBM, M4V or MKV, around 100 MB).
3. Choose Animate or Replace — Animate character keeps your image's background. Replace person drops your character into the clip's scene.
4. Pick 480p or 720p and generate — Describe the character's look if you like, set a seed to repeat a result, and download an MP4 with the clip's sound.

## In depth

### 3D-consistent pose, in plain words
Most motion engines read a flat 2D skeleton from each frame, which gets confused when a dancer turns away or crosses their arms. SCAIL's pose representation keeps depth, so the model knows which arm is in front and which way the body faces. The paper pairs this with 'full-context' pose injection: the model sees the whole motion sequence at once rather than frame by frame.

### Where that helps
It doesn't remove every artefact — very fast motion blur and props covering hands can still trip it up.
- Spins, turns and flips that confuse 2D skeletons.
- Stylised characters whose proportions differ from the dancer's.
- Scenes where limbs overlap or cross the body.

### SCAIL 2 in a real workflow
A typical run: test the first 15 seconds at 480p in Animate mode, compare the same seed in Replace mode, then render the full clip at 720p. If a fast spin wobbles, try the same clip on SteadyDancer; if a close-up face drifts, try Kling 3.0 Pro Motion.

## FAQ
Q: What does SCAIL stand for?
A: SCAIL stands for Studio-grade Character Animation via In-context Learning. The name comes from the research paper 'SCAIL: Towards Studio-Grade Character Animation via In-Context Learning of 3D-Consistent Pose Representations', accepted to the CVPR 2026 Findings track, whose authors released the model code and weights on GitHub under zai-org.

Q: Who made SCAIL?
A: SCAIL comes from zai-org, which open-sourced the SCAIL-Preview model, a 14B diffusion transformer, together with its pose extraction tools in December 2025 under the Apache 2.0 licence. On Kenerate AI, SCAIL 2 and SCAIL appear under WaveSpeed, the provider that serves them to the Motion Transfer tool.

Q: Is SCAIL 2 the default engine?
A: Yes. When you open Kenerate AI's Motion Transfer tool, SCAIL 2 is selected, and the 'Street dance' starter example uses it. It is the default because it covers the most cases at once: animate or replace modes, clips up to 2 minutes, 480p or 720p, and the clip's sound kept in the result.

Q: Does SCAIL keep the audio?
A: SCAIL 2 does: the driving clip's audio comes through in the result, so a dance stays on the beat and a talking clip keeps its speech. The first SCAIL engine returns silent video. Other engines that keep sound are Kling 3.0 and Kenerate Motion with a switch, LTX-2 Control and DreamActor v2.

Q: What is the difference between Animate and Replace in SCAIL 2?
A: Animate character moves the character in your image on your image's own background, following the clip's motion. Replace person keeps the clip's scene, camera and timing and swaps the person in it for your character. Use Animate for a clean character shot, and Replace when the original setting is the point.

Q: What happens if my clip is longer than 2 minutes?
A: SCAIL 2 uses the first 2 minutes and trims the rest; the job isn't refused. Since you pay per second of clip, it's worth trimming to the part you need before uploading. If a job fails for any reason, its credits are refunded automatically, so a test run costs nothing when it doesn't work.

Q: Can I use SCAIL on a photo of a real person?
A: Only with that person's permission. Use your own photos, characters you created, or people who agreed to it — never celebrities or strangers. SCAIL's Replace mode makes convincing swaps, so label the result as AI-made when you share it, and never use it to put someone into a scene they would object to.

## Related pages
- AI Motion Control: https://kenerateai.com/ai-motion-control
- Wan 2.2 Animate: https://kenerateai.com/wan-2-2-animate
- Replace a person in a video: https://kenerateai.com/replace-person-in-video
- AI character animation: https://kenerateai.com/ai-character-animation
- Make a photo dance: https://kenerateai.com/make-photo-dance

## Sources
- Kenerate Motion Transfer tool: https://kenerateai.com/app/tool/motion
- zai-org/SCAIL on GitHub: https://github.com/zai-org/SCAIL
- SCAIL paper (arXiv 2512.05905): https://arxiv.org/abs/2512.05905
- SCAIL project page: https://teal024.github.io/SCAIL/